NasoVoce Review: Nose-Pad Low-Audibility Speech Interface
DOIOriginal paper (ACM DOI) This review
This expert review explains NasoVoce’s nose-pad microphone and vibration sensing, and the conditions for private whispered input. Low-audibility speech is not fully silent speech; read the operating conditions below before treating it as hands-free, confidential AI interaction.
- What it is
- An expert review of NasoVoce: nose-pad dual-sensor (air + vibration) capture with multimodal enhancement for open-vocabulary low-audibility and whispered speech under noise, evaluated with ASR and perceptual quality metrics reported in the paper.
- Who it’s for
- wearable speech and SSI-adjacent researchers comparing nose-pad multimodal sensing to ultrasound or EPG routes who need a bounded expert take—and a clear next path into SilentSpeller, SottoVoce, or the silent-speech hub—before opening the paper.
- Verdict
- Strong deployment-oriented wearable speech reference for discreet low-audibility input; keep claims inside the paper’s sensor-fusion, Whisper-assisted enhancement, and evaluated noise settings, and do not equate it with zero-audio EPG/ultrasound silent spelling alone.
- Read next
- True value
- Demonstrates a feasible, socially acceptable, and noise-robust wearable speech interface design embracing discreet whispered speech input complemented by a novel hardware sensor fusion architecture and rigorous multi-modal enhancement and ASR evaluation, moving beyond small-vocabulary silent speech towards open-vocabulary AI conversation.
BibTeX
@misc{nasovoce,
title = {NasoVoce: A Nose-Mounted Low-Audibility Speech Interface for Always-Available Speech Interaction},
author = {Jun Rekimoto and Yu Nishimura and Bojian Yang},
year = {2026},
note = {CHI '26 / arXiv},
doi = {10.1145/3772318.3791397},
eprint = {2603.10324},
archivePrefix = {arXiv},
url = {https://doi.org/10.1145/3772318.3791397},
} Is NasoVoce fully silent, hands-free and private?
Not fully silent: NasoVoce captures normal and whispered speech, including airborne sound. Its microphone complements vibration sensing; it does not decode soundless mouth movements alone.
For private input, Section 3 and Figure 3 describe covering the mouth and nose with a hand to limit sound leakage and obscure lip movements. That privacy-oriented posture uses a hand, even though the sensors sit on glasses. It is different from claiming fully hands-free private input.
Section 5 evaluates recognition, reconstructed-audio quality and real-world recordings—not a bystander eavesdropping-distance guarantee. Noise robustness means recovering the user’s signal amid noise; it does not establish that others cannot hear the utterance. Check the intended setting separately if confidentiality matters. The links below separate the operating description from the evaluation.
Reading guidance
- Verdict
- full-text draft · priority high · confidence high
- Why it matters
- Demonstrates a feasible, socially acceptable, and noise-robust wearable speech interface design embracing discreet whispered speech input complemented by a novel hardware sensor fusion architecture and rigorous multi-modal enhancement and ASR evaluation, moving beyond small-vocabulary silent speech towards open-vocabulary AI conversation.
- What to trust
- Basis: full text + existing expert seed. Coverage: high. 4 evidence records back the review.
- What is weak
- Fusion model not fully streaming; whispered vibration signals remain weak limiting enhancement quality; performance under extreme noise favors vibration sensor input only at very low SNR. Evaluations used synthetic noise corruption for ASR and objective metrics; in-the-wild testing was qualitative with limited environments; no unseen words generalization tested. Fully streaming continuous operation, smartphone integration, adaptive sensor gating based on SNR, and calibration for physiological factors such as nasal patency remain future work. Targets low-audibility whispered speech, not fully silent speech without any acoustic leakage; assumes hand-covering mouth for privacy. Overclaim risk: medium.
- Read before
- SottoVoce: An Ultrasound Imaging-Based Silent Speech Interaction Using Deep Neural Networks
- Read next
- SSI archive
Axes
- Task
- speech-enhancement; speech-recognition for whispered and low-audibility speech to support always-available AI voice interaction
- Modality
- acoustic; vibration; multimodal
- Hardware
- MEMS microphone (Syntiant SPH0141LM4H-1) and MEMS vibration sensor (Syntiant V2S200D) integrated in smart glasses nose pads providing synchronized PDM output.
- Body site
- face; nose; oral-cavity
- Output
- speech-audio
- Vocabulary
- Open-vocabulary speech
- Metrics
- Recognition accuracy expressed in word error rate (WER) and character error rate (CER) under varying noise conditions; PESQ and STOI perceptual audio quality; MUSHRA subjective audio quality scores.
- Evaluation mode
- Quantitative ASR accuracy (WER, CER) on held-out data, objective perceptual quality metrics (PESQ, STOI), MUSHRA subjective ratings with 50 evaluators, and qualitative in-the-wild recordings in four real-world environments.
- Review confidence
- high
- Overclaim risk
- medium
Expert take
NasoVoce represents a serious, well-constructed system integrating a discreet smart-glasses nose-pad form factor housing a MEMS microphone and vibration sensor. By fusing these complementary modalities via a novel dual-input D-DCCRN enhancement model trained with Whisper-based knowledge distillation, it robustly captures low-volume and whispered speech. The extensive dataset of 104 hours from 45 participants, rigorous evaluation covering ASR metrics, objective speech quality evaluation, a large-scale subjective study, and real-world qualitative trials establish its practical feasibility. While the core method is validated, deployment aspects such as continuous streaming, adaptive sensor fusion depending on context, smartphone integration, and accommodation for physiological variability like nasal patency require further study and development.
True value
Demonstrates a feasible, socially acceptable, and noise-robust wearable speech interface design embracing discreet whispered speech input complemented by a novel hardware sensor fusion architecture and rigorous multi-modal enhancement and ASR evaluation, moving beyond small-vocabulary silent speech towards open-vocabulary AI conversation.
What changed
Canon before
Wearable silent speech and whispered speech interfaces typically struggled to balance wearability, vocabulary size, noise robustness, and social acceptability, often limited to small command sets or requiring obtrusive sensors.
Delta from canon
NasoVoce innovates by mounting a MEMS microphone and vibration sensor at the smart glasses nose pad, capturing complementary air- and skin-conducted signals for robust low-volume and whispered speech capture, combined with a dual-input enhancement model.
Position in field
A strong modern SSI-adjacent wearable speech paper that refocuses from fully silent recognition to discreet, robust AI voice interaction with open vocabulary.
Evidence
“ Lip-reading techniques have an extended • A whisper input mechanism by integrating a microphone and vocabulary size, but camera-based systems covering the face impose vibration sensor into nose pads; This design reduces ambient high wearability costs, limiting their suitability for daily use. noise for both normal and whispered speech while maintaining As a method related to silent speech, “whispered speech” has the smart glasses’ appearance and wearability. also been proposed. ”
author_claim · Abstract · confidence 1.00
“ Lip-reading techniques have an extended • A whisper input mechanism by integrating a microphone and vocabulary size, but camera-based systems covering the face impose vibration sensor into nose pads; This design reduces ambient high wearability costs, limiting their suitability for daily use. noise for both normal and whispered speech while maintaining As a method related to silent speech, “whispered speech” has the smart glasses’ appearance and wearability. also been proposed. ”
actual_novelty · 3 NasoVoce · confidence 1.00
“ 0 0 recorded MEMS vibration sensor (Vib) channel, microphone sig- 20 10 0 -10 20 10 0 -10 Signal to noise ratio [dB] Signal to noise ratio [dB] nals obtained by mixing the clean speech with various noises (Mic conditions), and signals enhanced from Vib+Mic by the proposed Figure 7: Speech recognition accuracy (WER, CER) for nor- audio-enhancement model. mal and whispered speech: MEMS microphone (Mic), MEMS vibration sensor (Vib), audio enhancement by D-DCCRN 5.1 ASR (WER /CER) Fig. ”
metric · 5 Evaluation · confidence 1.00
“ Lip-reading techniques have an extended • A whisper input mechanism by integrating a microphone and vocabulary size, but camera-based systems covering the face impose vibration sensor into nose pads; This design reduces ambient high wearability costs, limiting their suitability for daily use. noise for both normal and whispered speech while maintaining As a method related to silent speech, “whispered speech” has the smart glasses’ appearance and wearability. also been proposed. ”
limitation · 6 Discussions · confidence 1.00
Limits
Technical limits
Fusion model not fully streaming; whispered vibration signals remain weak limiting enhancement quality; performance under extreme noise favors vibration sensor input only at very low SNR.
Evaluation limits
Evaluations used synthetic noise corruption for ASR and objective metrics; in-the-wild testing was qualitative with limited environments; no unseen words generalization tested.
Deployment limits
Fully streaming continuous operation, smartphone integration, adaptive sensor gating based on SNR, and calibration for physiological factors such as nasal patency remain future work.
Scope limits
Targets low-audibility whispered speech, not fully silent speech without any acoustic leakage; assumes hand-covering mouth for privacy.